sebastienlevert/personal-toolkit

Address review feedback and merge conflicts on an existing PR end to end in any project. Use for "address PR feedback", "fix review comments", or "resolve PR conflicts". For authorized feedback fixes, assess the concern, ship validated code to the same PR branch, reply with commit/test evidence, then resolve only fully addressed review threads. Preserve unrelated work and PR state; never merge the PR. Plan-only and conflict-only requests remain scoped to that intent.

Evidence-backed review of pull requests, local changes, and follow-up fixes in any repository or technology stack. Use for "auto-review", "review this PR", "review my changes", "re-review", or an explicitly requested PR sweep. Applies architecture and implementation lenses plus change-specific checks, prioritizes compatibility and proportionate complexity, and consolidates actionable findings. Read-only by default; posting, approval, and recurring sweeps require an explicit request. Does not fix code or merge PRs.

Orchestrate existing development skills from issue triage/selection through implementation, review, feedback repair and PR readiness. Modes draft, ready and merge control final authority; draft is default. Invoke the owning skills rather than duplicating their logic. Explicit merge mode requires current, evidence-backed high confidence and normal repository checks/reviews; never self-approve, bypass protection or run an unbounded queue.

Deliver an authorized issue, feature, or bug fix end to end in any repository: refresh the actual base, use a dedicated sibling worktree and branch, implement the smallest complete change, use a brief test plan and lightweight readiness check, then push and open a draft PR. Reuse the same worktree and PR for follow-ups. Use for "implement this", "implement issue", or an explicitly requested implementation task. Never merge, ready a draft, overwrite unrelated work, or substitute a plan for delivery.

Explicitly initialize or refresh a selected folder or drive as a personal-toolkit workspace. Install the bundled AGENTS.md routing template while preserving existing instructions and local registry data. Use only when the user explicitly invokes init-toolkit; never infer initialization from ordinary repository work.

Investigate a bug, issue, failure, or unclear behavior in any project. Reproduce when authorized, trace the root cause, assess the impact, and return a bounded resolution and verification plan. Use for "investigate", "root cause", or "understand this failure". Advisory by default: no source fixes, GitHub posts, new dependencies, or automatic handoff to implementation.

Author or update an architecture decision record in any project using its own conventions. Use for "new ADR", "record this architectural decision", or "document the alternatives and tradeoffs". Separate proposed from accepted decisions; use a fresh-base task worktree and draft PR for persisted changes. No product-specific tooling, mandatory ADR gate, or automatic implementation.

Turn a product outcome into a grounded epic with priorities, scope boundaries, measurable completion criteria, and optionally approved child issues in any GitHub project. Use for "new epic", "create an epic", "epic proposal", or "break this epic into features". Research existing work and preserve open decisions; do not implement features or invent customer demand.

Research, deduplicate, draft, or create an actionable GitHub issue in any project using its own conventions and the gh CLI. Use for "new issue", "file an issue", "create an issue", or "draft an issue". Preserve the original request, distinguish evidence from assumptions, and define observable acceptance. Creates only the requested issue; does not implement, create a spec, or open a PR.

Research and author or update a focused specification in any project, following its contracts, document conventions, and acceptance criteria. Use for "new spec", "write a spec", "specify this feature", or "update this specification". Persisted repository changes use a dedicated worktree from a refreshed base and finish with validation and a draft PR. Draft-only requests remain in chat.

Create or refine decision-ready strategy and architecture papers, using evidence, a clear thesis, alternatives, tradeoffs, and a concrete ask. Use for "write a strategy paper", "decision paper", or "rich HTML paper". Produce editable Markdown and, when requested, clean self-contained HTML with light/dark mode, a floating TOC, reading-time estimate, and SVG-only visuals. Ask only material missing questions; do not impose a fixed interview or publish automatically.

Publish reviewed findings as inline GitHub pull-request comments at the exact relevant diff lines using the gh CLI. Use for "post PR comments", "post this review", "post the comments inline", or "send these findings to the PR". Verify the current revision, preserve finding IDs, deduplicate delivery and verify published threads. Posting is explicit; never approve, request changes, resolve threads, edit source or merge without separate authorization.

Assess a PR or prospective draft for review readiness using a lightweight check of scope, diff, tests, project rules and available CI evidence. Use for "prepare PR", "is this PR ready", or "check readiness". Assessment is read-only; fixes and marking ready need explicit authorization. Integrated into implement as one final pass that reuses current validation and leaves its PR a draft.

Produce decision-oriented market, competitive, customer, or technical research for any project. Use for "research this", "competitive brief", "compare these options", or "technical landscape". Verify sources, dates and comparison conditions; separate facts, vendor claims, inference and unknowns. Return a concise recommendation and evidence, with optional Markdown/HTML handoff to paper-strategist. No automatic publication, private-data export or implementation.

Clone or explicitly migrate a repository into a bare Git container with .bare metadata, a default-branch checkout such as main, and task worktrees beside it inside the container. Use for "setup repo", "clone with worktrees", or "convert to bare worktree layout". Preserve refs, local work, remotes and configured source paths; never overwrite an existing checkout or delete repository roots.

Derive a concise, risk-based test plan from a task, local changes, or a PR in any project. Cover intended behavior, important edge cases, compatibility, and existing validation commands. Use for "test plan" or "what should I test". Standalone use is read-only and runs no tests. Also supplies the lightweight planning step inside implement without requiring a second review or artifact.

Triage a bounded issue backlog or review feedback against current evidence and project intent. Backlog mode selects the next eligible issue and explains blockers; feedback mode returns fix, clarify, defer, no-change or information dispositions. Use for "select next issue", "triage backlog", or "triage review comments". Advisory by default; corrections and publication require explicit scope and authorization.